Many clients like to use the Web to discover product details in the form of online reviews. The reviews are provided by other clients and specialists. Recommender systems provide an important response to the information overload problem as it presents users more practical and personalized information facilities. Collaborative filtering methods are vital component in recommender systems as they generate high-quality recommendations by influencing the likings of society of similar users. The collaborative filtering method has assumption that people having same tastes choose the same items. The conventional collaborative filtering system has drawbacks as sparse data problem & lack of scalability. A new recommender system is required to deal with the sparse data problem & produce high quality recommendations in large scale mobile environment. MapReduce is a programming model which is widely used for large-scale data analysis. The described algorithm of recommendation mechanism for mobile commerce is user based collaborative filtering using MapReduce which reduces scalability problem in conventional CF system. One of the essential operations for the data analysis is join operation. But MapReduce is not very competent to execute the join operation as it always uses all records in the datasets where only small fraction of datasets are applicable for the join operation. This problem can be reduced by applying bloomjoin algorithm. The bloom filters are constructed and used to filter out redundant intermediate records. The proposed algorithm using bloom filter will reduce the number of intermediate results and will improve the join performance.

1. International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.3, No.6, November 2013
RECOMMENDATION SYSTEM USING BLOOM
FILTER IN MAPREDUCE
Reena Pagare and Anita Shinde
Department of Computer Engineering, Pune University
M. I. T. College Of Engineering Pune India
ABSTRACT
Many clients like to use the Web to discover product details in the form of online reviews. The reviews are
provided by other clients and specialists. Recommender systems provide an important response to the
information overload problem as it presents users more practical and personalized information facilities.
Collaborative filtering methods are vital component in recommender systems as they generate high-quality
recommendations by influencing the likings of society of similar users. The collaborative filtering method
has assumption that people having same tastes choose the same items. The conventional collaborative
filtering system has drawbacks as sparse data problem & lack of scalability. A new recommender system is
required to deal with the sparse data problem & produce high quality recommendations in large scale
mobile environment. MapReduce is a programming model which is widely used for large-scale data
analysis. The described algorithm of recommendation mechanism for mobile commerce is user based
collaborative filtering using MapReduce which reduces scalability problem in conventional CF system.
One of the essential operations for the data analysis is join operation. But MapReduce is not very
competent to execute the join operation as it always uses all records in the datasets where only small
fraction of datasets are applicable for the join operation. This problem can be reduced by applying
bloomjoin algorithm. The bloom filters are constructed and used to filter out redundant intermediate
records. The proposed algorithm using bloom filter will reduce the number of intermediate results and will
improve the join performance.
KEYWORDS
Collaborative Filtering, MapReduce, Hadoop, Recommender System, Recommender Algorithm, Bloom
filter
1. INTRODUCTION
Since the amount of information in e-commerce & mobile commerce increases, there is need to
filter irrelevant information & discover helpful contents & reliable sources. Recommender
system is the standard tool which gives advice about items, products, information or services
users might be fond of. Recommendation systems generate a ranked list of items on which a user
might be interested. Recommendation systems are constructed for movies, books, communities,
news, articles etc. They are intelligent applications to assist users in a decision-making process
where they want to choose one item amongst a potentially overwhelming set of alternative
products or services [1]. Recommender systems are personalized information filtering technology
used to either predict whether a particular user will like a particular item or to identify a set of N
items that will be of interest to a certain user. It is not necessary that a review is equally useful to
all users. The review system allows users to evaluate a review’s support by giving a score that
ranges from “not helpful” to “most helpful”. If a particular review is read by all users & found
helpful then it can be assumed that new user might appreciate it. Controversial reviews are the
reviews that have a variety of conflicting rating (ranking). Controversial review has both
DOI : 10.5121/ijdkp.2013.3608
127

2. International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.3, No.6, November 2013
passionate followers and motivated enemy without clear majority in either group. The
Recommender System uses information from user profile & interaction to tell possible items of
concern. It is useful to approximate the degree to which specific user will like a specific product.
The Recommender systems are useful in predicting the helpfulness of controversial reviews [2].
Recommender systems are a powerful new technology for extracting additional value for a
business from its user databases. These systems help users to find items they want to buy from a
industry. Recommender systems benefit users by helping them to discover items they like. They
help the business by generating more sales. Recommender systems are fastly becoming a crucial
tool in E-commerce on the Web [3].
Most commonly used algorithm for personalized recommendation in commercial
recommendation system is Collaborative Filtering (CF). The main problem of CF is its scalability
means the calculation cost of CF would be high if the dataset is very large. MapReduce has been
widely used for large scale calculation job. It is used to process massive amount of data in a
reasonable amount of time. The main advantages of MapReduce are straightforward
programming interface and high scalability along with refined failure management. But
MapReduce has some restrictions while doing join operation on large scale datasets. The main
difficulty in join processing in MapReduce is that complete dataset should be processed and sent
to nodes. This creates bottleneck in performance particularly when only small fraction of data is
applicable for the join. In this paper, we implement bloom join algorithm on the Hadoop, an
open-source execution of MapReduce framework, to enhance the join performance. With this
method, we can avoid redundant records and decrease communication overhead.
The rest of this paper is organized as follows. Section 2 explains related work in mobile
commerce, MapReduce, joins in MapReduce and bloom filter. Section 3 introduces User- based
Collaborative Filtering recommendation algorithm on Hadoop platform. Section 4 describes the
proposed architecture. Section 5 shows experiment results and examination. Finally Section 6
concludes the paper.
2. RELATED WORK
2.1 Mobile commerce
In [4], mobile commerce overview, its business applications and technical environment for
mobile commerce are explained. Different business needs relevant to mobile commerce are
described. It also helps business to define benefits from mobile commerce. It guides about how to
use mobile devices to connect people with information about products and to improve association
between trading partners in the supply chain and between businesses and customers.
[5] It gives an overview of the basics about m-commerce and e commerce and connection
between them. It helps the business managers who are not from IT background to realize the key
elements & fundamental problems of m-commerce. It also explains how to determine impact of
m-commerce on current and future businesses & to discover new business scenarios. It assists
businesses to define benefits derived from mobile commerce and describes the types of mobile
commerce applications.
M-commerce history, present information and essential metrics to evaluate usefulness of MCommerce are described. It focuses on easy m-commerce ideas to make mobile sales and on how
to build website mobile friendly [6].
128

3. International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.3, No.6, November 2013
2.2 MapReduce
MapReduce is a programming model for expressing distributed computations on huge amounts of
data and an execution framework for large-scale data processing on clusters of product servers. It
was initially developed by Google and built on well-known principles in parallel and distributed
processing.
A MapReduce program consists of two functions: map and reduce. The map function accepts a
set of records from input files in the form of simple key-value pairs and constructs a set of
intermediate key-value pairs. The values in the intermediate pairs are automatically collected by
key and sent to the reduce function. The grouping consists of sort and merge processes. The
reduce function accepts an intermediate key and a set of values related to the key, and finally it
generates output key-value pairs [7].
Figure1. Execution Overview of MapReduce
A MapReduce cluster consists of one master node and a number of worker nodes[8]. The master
at regular intervals communicates with each worker to test their status and manage their
performance. When a MapReduce job is acceptted, the master divides the input data and
generates map tasks for the input splits, and reduce tasks. The master allocates each task to idle
workers. A map worker accepts the input split and implements map task given by the user. A
reduce worker reads the intermediate pairs from all map workers and accomplish reduce task.
When all tasks are complete, the MapReduce job is finished.
2.3 Joins in MapReduce
MapReduce’s Join algorithms are generally categorized into two classes: map-side joins and
reduce-side joins [9]. Map-side join algorithms are more well-organized than reduce-side joins as
only final result of the join is constructed by them in map phase. For Map-Merge join , two input
datasets need to be separated and sorted using join key. Whereas reduce-side join algorithms are
more common, still they are not competent as the whole input records need to be transferred from
map workers to reduce workers[10].
2.4 Bloom filter
A Bloom filter [11] is a probabilistic data structure which is used to check whether an element is
a present in a set. It consists of an m-bits array and k self-sufficient hash functions. All bits in the
array are initialised to 0. When an element is included into the array, the element is hashed k
times using k hash functions, and the location in the array equivalent to the hash values are set to
1. In order to check membership of an element, all bits of its k hash positions of the array are
examined. If values in all bits are 1, we can say that the element is a member of set.
Bloom-join [12] is a join algorithm which uses the Bloom filter. It makes use of Bloom filter to
filter out tuples that are not matched in a join. For example, there are two relations X(a,b) and
Y(a,c) situated at location 1 and location 2 respectively. For joining these two relations, Bloomjoin creates a Bloom filter with the join key of relation X. Then it forwards the filter to location 2.
At location 2, the algorithm scans X and forwards only tuples which are set in the received Bloom
filter to location 1. At last, a join of filtered X and Y is executed.
129

4. International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.3, No.6, November 2013
3. USER BASED COLLABORATIVE FILTERING RECOMMENDATION
ALGORITHM ON HADOOP
3.1 Collaborative Filtering
Collaborative filtering algorithm is commonly used recommendation method in business-related
recommendation system which improves the performance recommendation process. The main
drawback of collaborative filtering is its scalability means when the size dataset is very huge, the
cost of calculation would be very high. Cloud computing tries to solve this problem of high
calculation task [1]. Collaborative filtering algorithm has assumptions: People have similar
likings & concern which are steady. It is possible to guess their option from their past
preferences. Collaborative filtering algorithms acquire users’ history summary in the form of a
ratings matrix consisting of rating given by user to every item. The next step is to compute the
similarity between users & discover their nearby neighbours. The most commonly used similarity
measure is the Pearson correlation coefficient which is standard for CF.
Where rx is rating of user x on item s and ry is rating of user y on item s, Sxy indicates the items
that user x and y co-evaluated.
The final step is to determine ratings of items which is average of the ratings by the neighbours
The CF algorithm requires rigorous computing and computer resources. The calculation process
would be very longer if the dataset is very huge.
The job in collaborative filtering is to guess the usefulness of product to a particular user which is
based on a database of user votes. Collaborative filtering algorithms guess ranking of a target
item for target user with help of grouping of the ranking of the neighbours (similar users) that are
known to item under consideration.
3.2 MapReduce and Collaborative Filtering
This part describes how Collaborative Filtering Algorithm can be implemented within the
MapReduce framework. It is difficult to directly use MapReduce model in computation process
of Collaborative Filtering algorithm. The recommendation process for each user is summarized in
the Map function i.e. while making recommendation, we will save user ID in text files which
serves as input to the Map function. The MapReduce framework defines few mapper to handle
the user ID files. The algorithm is partitioned into three stages as Data segmenting stage, Map
stage and Reduce stage.
In Data segmenting stage, the user ID is separated into various files in which every row has a user
ID. These files serve as input to the Map stage. This stage needs to fulfil two rules. The first rule
is that instead of repeatedly define the mapper, the maximum amount of the execution time
should be used up in calculation procedure. The second rule is that each mapper should have
same end time.
130

5. International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.3, No.6, November 2013
In Map stage, the Hadoop framework decides whether to define mapper to handle the user ID
files. The Hadoop framework defines a new mapper, if sufficient resources are available. Initially
the ratings matrix between the consumers and products is constructed by setup function of
mapper. Then it gets the line number as the input key and equivalent user ID as value. Then it
computes similarity between the user and other users. The last step is to determine the user’s
nearest neighbours based on similarity values and compute his expected rating on products.
These ratings are sorted and stored in recommend list. The user Id and its associated recommend
list is given to the Reduce phase.
Figure 2. Collaborative Filtering Using MapReduce
In Reduce stage, the Hadoop platform automatically creates some reducers which collect the user
ID and its associated recommend list, sort them as per their user ID and then forward to the
Hadoop Distributed File System(HDFS).
4. PROPOSED ARCHITECTURE
This section illustrates the overall architecture of our implementation. Hadoop is used in our
approach, an open-source implementation of the MapReduce framework. In Hadoop, we have
master node and worker node. The master node is called jobtracker and the worker node is called
tasktracker.
4.1 Execution Overview
Figure 2 shows the overall execution flow of join function on the datasets in our implementation.
When the user login to the system by entering username and password, the following sequence of
steps is carried out.
1.
Job submission: When user does login, the calculation of recommendations for active
user starts. The users in datasets are getting divided amongst the available tasktracker.
Tasktracker contains all the essential information required for map phase.
2.
First Map phase: The jobtracker gives the map tasks to idle tasktrackers. A map
tasktracker accepts the input split for the task, transfers it to key/value pairs, and then
executes the map function for the input pairs.
131

6. International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.3, No.6, November 2013
3.
Local filter construction: Bloom filter is constructed on the key. This filter is called
local as they are created only the intermediate results in a tasktracker. Here we get the
map output of local filter.
4.
Global filter merging: All map output of local filter is given to jobtracker which will
construct the global filter.
5.
Second Map Phase: Tasktrackers will execute the given task with received global filter.
The Redundant records are filtered out with the help of global filters.
6.
Reduce phase: A reduce tasktracker examine the resultant intermediate pairs from all
map tasktrackers. It arranges the all intermediate pairs and executes the reduce function.
At the end, output results are generated from the reduce function.
Figure3. Execution Overview
5. EXPERIMENTAL RESULTS AND ANALYSIS
In this section, we show our experiment result. We implement User based Collaborative Filtering
algorithm on Hadoop platform. Five computers Hadoop cluster was constructed. All experiments
were run on this cluster which has one jobtracker (NameNode) and remaining four tasktrackers
(DataNodes). Each computer has 4GB RAM & INTEL 2.5GHZ CPU and operating system
Ubuntu 10.10. The software used for the experiments includes Hadoop MapReduce framework,
Java JDK 1.6. The additional hardware required are the Mobile device (Android 3.0 & above) and
switch. In experiments, the Netflix data set is used. The Netflix dataset consists of around 17770
movies and 4, 80,189 users. The main aim of our CF algorithm is to compare the runtime between
standalone and Hadoop platform so accuracy is not focussed. We have taken 4 copies of subdatasets as our experimental datasets which include 100 users, 200users, 500 users and 1000
users. Similarly, DataNode number is considered as 2 nodes, 3 nodes and 5 nodes.
132

7. International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.3, No.6, November 2013
The local filter is constructed on the time interval as months 3, 6,9,12 etc. Then we get the map
output of the time interval as we will consider only those reviews of given time interval from
specified date. All map output of the local filters is given to job tracker which will construct
global filter which will be derived from the movie i.e. How many rating for considered movies.
Finally we get reduce operation output. The number of intermediate /output records is constructed
from the interval in months (i.e. 3, 6,9,12 etc), map output from reviews and map output from
movies. Finally join of reviews and movies will give the reduce output. In this implementation,
map phase is terminated early as the intermediate results are decreased by the bloom filters.
Similarly the reduce phase is also terminated early as number of intermediate records are
decreased, so the number of input records to work with in reduce function are decreased. It is
essential to decide the suitable dimension of the Bloom filter. If the dimension of Bloom filter is
small, it will be unable to filter out redundant records. Similarly, if the dimension of filters is
bulky, it will create large overhead to design and communicate filters.
6. CONCLUSIONS
Bloom filters used in MapReduce will help to reduce the intermediate results in map phase which
in turn speed up the overall process of recommendation. This paper shows how MapReduce can
be used to parallelize Collaborative Filtering. It also presents an architecture to enhance the join
performance using Bloom filters in the MapReduce framework. We propose to apply this
concept to Recommender System for Mobile Commerce.
In future work, we will expand proposed architecture to handle multi-way joins and to determine
the appropriate size of Bloom filters.
REFERENCES
[1]
Zhi-Dan Zhao, Ming-Sheng Shang. User-Based Collaborative-Filtering Recommendation
Algorithmson Hadoop[C]. In Proceedings of the Third International Conference on Knowledge
Discovery and Data Mining, (2010) 478 – 481.
[2] Adomavicius G., Tuzhilin A., Toward the next generation of recommender systems: A survey of the
state-of-the-art and possible extensions. IEEE Trans. on Knowledge and Data Engineering, 2005,
17(6): 734-749
[3] Reena Pagare and Anita Shinde Article : A Study of Recommender System Techniques. International
Journal of Computer Applications 47(16):1-4, June 2012. Published by Foundation of Computer
Science, New York, USA.
[4] Mobile Commerce: opportunities and challenges A GS1 Mobile Com White Paper February 2008
Edition
[5] Asghar Afshar Jahanshahi, Alireza Mirzaie, Amin Asadollahi MOBILE COMMERCE BEYOND
ELECTRONIC COMMERCE: ISSUE AND CHALLENGES Asian Journal of Business and
Management SciencesISSN: 2047-2528 Vol. 1 No. 2 [119-129] www.ajbms.org
[6] MOBILE COMMERCE The Future Is Here 3dcart.com/whitepapers ©2010, 3dcart
[7] J. Dean and S. Ghemawat. Mapreduce: Simplified data processing on large clusters. In Proceedings of
the 6th USENIX Symposium on Opearting Systems Design & Implementation (OSDI), pages 137–
150, 2004.
[8] Taewhi Lee, Kisung Kim, and Hyoung-Joo Kim Join Processing Using Bloom Filter in MapReduce
RACS’12 October 23-26, 2012, San Antonio, TX, USA. 2012 ACM 978-1-4503-1492-3/12/10
doi>10.1145/2401603.2401626
[9] K.-H. Lee, Y.-J. Lee, H. Choi, Y. D. Chung, and B. Moon. Parallel data processing with mapreduce:
A survey. ACM SIGMOD Record, 40(4):11–20, 2011.
[10] S. Blanas, J. M. Patel, V. Ercegovac, J. Rao, E. J.Shekita, and Y. Tian. A comparison of join
algorithms for log processing. In Proceedings of the 2010 ACM SIGMOD International Conference
on Management of Data (SIGMOD ’10), pages 975–986, 2010.
133

8. International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.3, No.6, November 2013
[11] B. H. Bloom. Space/time trade-offs in hash coding with allowable errors. Communications of the
ACM (CACM), 13(7):422–426, 1970.
[12] L. F. Mackert and G. M. Lohman. R* optimizer validation and performance evaluation for distributed
queries. In Proceedings of the 12th International Conference on Very Large Data Bases (VLDB),
pages 149–159, 1986.
[13] Zan Huang, Daniel Zeng and Hsinchun Chen A Comparison of Collaborative-Filtering
Recommendation Algorithms for E-commerce 1541-1672/07/$25.00 © 2007 IEEE
AUTHORS
Mrs Reena Pagare has done her graduation and post graduation from Pune
University, India in 2001 and 2009 respectively. She is currently working with Maeer's
Maharashtra Institute of Technology ,Computer Department, Pune, India.She has 10
yrs of teaching experience. She has published numbers of national and international
papers. Her research interest are Recommender Systems, Algorithms.
Anita Shinde has received her BE in 2003 from Pune University and persuing ME
from Pune University. She is currently working with MM College of Engineering,
Computer Department, Pune, India. She is life member of ISTE. She has published
numbers of national and international papers.Her research interest includes
Recommender Systems, data mining and information retrieval.
134